Iterating on files confuses Assistant

Phemist

I often try to iterate with the assistant on a file that I initially uploaded. It goes something like this:

Turn 1 - Me: Upload file + question/request assistant to please fix X
Turn 2 - Assistant: This is the fix (puts updated file in context)
Turn 3 - Me: Thanks, I changed it. Another question assuming the new state of the file (assuming the assistant uses the file from context)
Turn 4 - Assistant: Ah, let me check the uploaded file! Oh you didn't change anything at all.

And then things go sideways from there. I tried to diagnose the reason with the assistant. The first 4 turns show the issue cleanly:

https://kagi.com/assistant/a63fb945-5f44-42fd-9af0-64cb4e52ee06

The latter turns are me trying to diagnose the problem. It seems the system context prompt is injected every turn, with no clear indication that it is an old and stale file from a previous turn.

I would like the Assistant to recognize that the uploaded files are now stale, as these were included in a previous turn. I would then like it to recognize that it should be using information from its own context.

Rehan

I'm a bit confused as to what exactly the issue is. You only upload one file, correct? Are you expecting it to read a different file?

Phemist

I am uploading 1 file. I ask the Assistant to change it in some way, confirm my changes on my end, and then ask it something related to the shared understanding that we improved the file I originally uploaded.

Rather than seeing that the Assistant's suggested changes happened after my file upload, and correctly thus answering e.g. MIT license, it re-reads the original file upload, forgets that it actually suggested an update, and answers Mozilla license.

Please contrast this with a local run of Qwen3.6-35b, see image:

Qwen3.6 does correctly notice that the moz.build file is stale and that considering I updated it, the new license is now MIT.

I would expect the Kagi Assistant to ALSO know this, otherwise these iterative improvements of what was originally a file (which everything is now with recent changes, even a large copy-paste of text) become incredibly tiresome. This is clearly a regression